Continuous-AdvTrain

This is a repository for the adversarial training code of our paper https://arxiv.org/abs/2405.15589.

Models

Our models can be found on Huggingface:

Phi-CAT

Phi-CAPO

Zephyr-CAT

Llama-2-CAT

Zephyr-CAPO refused a lot of harmless requests. We will upload a new model soon with a more sensible robustness utility trade-off

Installation

1. Clone this repository with git clone [email protected]:sophie-xhonneux/Continuous-AdvTrain.git
2. Install the requirements pip install -r requirements.txt

Running the code

1. Create a config in config/path (see example_path.yaml)
2. Run the code with python src/run_experiments.py --config-name=adv_train_ul path=example_path

You can also run the IPO experiments by replacing adv_train_ul with adv_train_ipo. Moreover, hydra allows you to override any hyperparameters from the commandline (e.g. add adversarial.eps=0.075) or you can create a new config file under the config folder. See the paper for the exact hyperparameters.

Data

The data is in the data folder is from the Harmbench repository (https://github.com/centerforaisafety/HarmBench) with the exception of a couple files we created as part of this paper.

Citation

If you used this code, please cite our paper:

@misc{xhonneux2024efficient,
      title={Efficient Adversarial Training in LLMs with Continuous Attacks}, 
      author={Sophie Xhonneux and Alessandro Sordoni and Stephan Günnemann and Gauthier Gidel and Leo Schwinn},
      year={2024},
      eprint={2405.15589},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}